Abstract 323 – Building a biometric: Leveraging pose-aware imagery to develop a robust photo-identification for an unmarked species (Ursus arctos)

Beth Rosenberg, Mu Zhou, Nathan Wolf, Bradley P. Harris, Alexander Mathis

The importance of distinguishing between individuals of a species is critical for answering a
range of biological and ecological questions. Despite current success in identifying individuals
with patterns or intra-species markings, numerous species exist for which the use of visual
individual identification is restricted or prevented by a lack of obvious uniqueness (e.g. a non-
patterned species in which a readily observable individual identifier is not apparent). Alaskan
coastal brown bears (Ursus arctos) are an example of a species for which the development of
successful and practicable visual individual identification methods is hindered by a lack of
obvious unique or permanent biometric traits. In addition to lacking intra-species markings,
individuals undergo weight gain, fur shed, and scarring which alters appearance and makes
identifying a non-patterned brown bear even more challenging. Robust individual identification
of this species requires the development of a novel technique or the building of a new
biometric. In this study, we take into consideration the physiology, morphology, and behavior
of the species to place conditions on an image dataset, mainly using a pose-aware framework.
We formulate a conditional neural network to provide “permanence” through a multiplicity of
image comparisons, utilizing images and postures that might otherwise be unusable. The
images required for individual brown bear identification are not limited to any one pose or
view, and can handle both occlusions, as well as unknown individuals. Leveraging a large
number of known individual brown bears over time in a unique data set, we demonstrate that
an individual ID for brown bears can be successfully predicted not only over time, but also
across the landscape at locations other than where a bear was first observed or where the
dataset was compiled.